MesoSplats: Texture Synthesis With Gaussian Splatting

Area of Science:

• Computer Graphics
• 3D Rendering
• Computational Geometry

Background:

• 2D texture mapping is standard but insufficient for complex volumetric meso-structures.
• Existing methods for meso-structure textures lack detail and real-time performance.

Purpose of the Study:

• To introduce a novel neural implicit method for high-fidelity texture extraction and synthesis of volumetric meso-structures.
• To enable real-time rendering of complex 3D assets with detailed textures.

Main Methods:

• Developed MesoSplats, a hybrid mesh-Gaussian representation leveraging 3D Gaussian splatting.
• Employed Consistency Tuning for refining local implicit texture features.
• Integrated tileability-aware patch matching and latent space smoothness regularization for synthesis quality.

Main Results:

• Successfully extracted and synthesized volumetric meso-structure textures from multi-view images.
• Achieved high-fidelity reconstruction and real-time rendering capabilities.
• Demonstrated superior performance in quantitative and qualitative experiments compared to existing methods.

Conclusions:

• MesoSplats effectively addresses limitations of 2D texture mapping for complex 3D assets.
• The proposed method enables realistic and efficient rendering of volumetric textures.
• Offers a promising direction for advanced 3D content creation and digital asset representation.